{"id":"d8c884d7-2a96-4736-b069-32d5fa82d0b5","arxiv_id":"2501.05399","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"YOLOv7 detects knife-handling hazards in a self-captured kitchen video dataset, reaching mAP50-95 0.7879 at epoch 31, but performs poorly on the blade-contact hazard class.","lead":"A single-author study trains YOLOv7 to spot kitchen knife hazards, reporting strong average precision but weak detection of blade-hand contact. The work is a small application study of an existing detector, with no released data or code.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported mAP50-95 is likely inflated by temporal autocorrelation and the absence of a held-out test set, so the headline numbers do not yet establish real-world detection performance.","rationale":"The reader's weakest assumption—single-video, single-annotator data—is real, but the more decisive issue is internal: the paper's headline numbers come from an unspecified validation procedure on frames sampled from the same video used for training (Sections 3.1 and 4). Consecutive video frames are temporally correlated, so the effective independent sample size is far below 6,004, and selecting the epoch with the best validation mAP on same-scene frames can overstate generalization. This is not a disagreement with community consensus; it is a standard correctness risk in video-based detection evaluation, and a concrete temporal-split test would settle it. Independently, the paper's own Section 4 reports hazard2 AP of 0.290, which undercuts the abstract's 'accurately detect knife-related hazards' phrasing; this is an internal inconsistency, not merely an external generalization issue. I therefore agree with the conditional verdict but place more weight on the missing test set and temporal leakage than on annotation expertise. The YOLOv7 methodology itself is standard, and the PR-curve and confusion-matrix figures provide some transparency, but they do not replace a held-out evaluation.","tokens_in":12185,"tokens_out":3900,"duration_ms":37261,"concrete_test":"Retrain the model with a strict temporal split of the single video: use the first 4,200 frames for training, the next 1,200 for validation, and the final 604 frames as a held-out test set, and report mAP50-95 on that held-out test. Then run the same trained model on a separate video recorded in a different kitchen session. If the held-out mAP50-95 is substantially below 0.7879, or hazard2 AP remains near 0.290, the reported numbers are validation-set artifacts rather than generalization results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—mAP50-95 of 0.7879, precision 0.9063, recall 0.7503 at epoch 31—rests on validation curves computed during training (§4). The paper never specifies how the 6,004 frames from Section 3.1 were split into training, validation, and test sets. Because these frames are consecutive samples from one continuous iPhone video, adjacent frames are near-duplicates with highly correlated content, pose, and lighting. If validation frames are drawn from the same video used for training, and the model is selected by best validation mAP, the evaluation is effectively self-evaluation on the training distribution; the numbers do not measure generalization to other kitchens, subjects, or camera angles. A second, internal inconsistency compounds this: Section 4 reports hazard2 AP of only 0.290 and a peak F1 of 0.75, contradicting the abstract's implication that both hazards are accurately detected. Even taking the validation numbers at face value, the most safety-relevant class—blade contact—is weak. Both issues must be resolved before the headline claim is accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a YOLOv7 object detector on 6,004 frames extracted from a single iPhone-recorded kitchen video, annotated into six classes (cutting board, hand, vegetable, knife, hazard 1 curled fingers, hazard 2 hand touching blade). It describes standard data augmentations, summarizes the YOLOv7 architecture, and reports training/validation curves: precision, recall, mAP50, per-class AP, F1-confidence, and a confusion matrix. The abstract claims best performance at epoch 31 with mAP50-95 of 0.7879, precision of 0.9063, and recall of 0.7503, and concludes that YOLOv7 can accurately detect knife-related hazards.","tokens_in":12548,"tokens_out":3739,"duration_ms":34366,"significance":"If the headline numbers were supported by a properly held-out evaluation, this would be a modest empirical datapoint for YOLOv7 in a niche safety domain, and the paper's transparent reporting of per-class weaknesses (especially hazard 2 AP of 0.290) is a strength. However, as presented, all quantitative claims are derived from validation curves on a single video with no independent test set, so the results do not currently establish generalization to other kitchens, users, or lighting conditions. The significance for real-world deployment therefore remains unsubstantiated, although the underlying idea is worth testing with a more rigorous evaluation protocol.","major_comments":[{"comment":"The headline metrics in the abstract (mAP50-95=0.7879, precision=0.9063, recall=0.7503 at epoch 31) do not appear anywhere in Section 4; the results section instead reports mAP50=0.821, per-class AP values, and a peak F1 of 0.75. The paper must either add the curves or tables that produce the abstract numbers or correct the abstract. More importantly, no training/validation/test split is described in Section 3.1, and all reported curves appear to be computed on the validation fold used for early stopping and threshold selection. Because the 6,004 frames are consecutive samples from one continuous video, adjacent frames are near-duplicates, so the reported numbers are effectively self-evaluation on the training distribution and do not measure generalization. Please add an explicit data split, ideally with a temporally separated test set or a separate video, and report metrics on that test set, or reframe all claims as validation performance with explicit caveats.","section":"Abstract and Section 4"},{"comment":"Hazard 2 (hand touching blade) achieves an AP of only 0.290 and is described as 'struggling' in the PR and F1 curves, yet the abstract concludes that YOLOv7 can 'accurately detect knife-related hazards.' Since hazard 2 is one of only two safety hazards the paper is designed to detect, this per-class weakness undercuts the headline claim regardless of mean metrics. Please report per-class results in the abstract and conclusion, or adjust the claim to state that only hazard 1 is detected well.","section":"Section 4, Figures 7 and 9"},{"comment":"The dataset consists of 6,004 frames from a single iPhone video, labeled by one person without reported expert validation or inter-annotator agreement. This single-annotator, single-video design is the sole source of ground truth, so the model's measured performance is bounded by that annotation and cannot support broad claims about kitchen safety. Please state the annotator's background, provide a labeling protocol or sample annotations, and explicitly discuss how the single-video/single-annotator design limits the conclusions.","section":"Section 3.1"}],"minor_comments":[{"comment":"Equation (2) is incomplete: the square bracket is never closed and the right-hand side has no closing bracket; please provide the complete expression and define all variables.","section":"Section 3.2.2"},{"comment":"There are typographical errors: 'Grayscle' should be 'Grayscale' in the Figure 2 caption, and 'he architecture' should be 'The architecture' in Section 3.3.","section":"Figure 2 caption and Section 3.3"},{"comment":"The text repeatedly refers to 'accuracy improvements' and 'high accuracy,' but no accuracy metric is defined; the reported measures are precision, recall, mAP, and F1. Please use the appropriate terminology.","section":"Section 4"},{"comment":"In the confusion-matrix description, 'knife' is reported as 0.85 and misclassified as 'vegetable' with 0.32, yet these values do not sum to 1; clarify whether rows are normalized and whether background cells are included in the normalization.","section":"Section 4, Figure 10"},{"comment":"The reference list contains duplicate entries (for example, refs [9] and [28] are the same paper, and refs [20] and [52] are the same) and mixes peer-reviewed sources with blog posts; please unify and curate the bibliography.","section":"References"},{"comment":"The augmentation section lists many transforms but does not state the actual probabilities or strength values used in the final training pipeline; please report the concrete augmentation configuration used for the reported experiments.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a dense self-citation pattern (refs [44]-[47]) referencing the author's own previous work on the same kitchen-safety dataset; while not disqualifying, the editor may wish to ask the authors to clarify the novelty relative to the earlier comparative study in [44]. In addition, no code or dataset is provided, and the journal should consider whether a single-video validation study with no independent test set meets its bar for empirical contributions; a major revision that adds a proper held-out evaluation and reconciles the abstract numbers with the results section would address this concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a small paper: a single-author evaluation of YOLOv7 on a self-collected six-class kitchen knife dataset. The new thing is the dataset — the specific hazard classes (curled fingers, hand touching blade) weren't in the author's earlier YOLOv5/v8/v10 comparison, and I don't see another YOLOv7 kitchen-safety study in the citations. The paper does a decent job describing the augmentation pipeline and it is transparent in Section 4 that hazard 2 is weak (AP 0.290, peak F1 0.75). That honesty in the results is worth something.\n\nThe soft spots are real, though. The abstract's headline numbers — mAP50-95 0.7879, precision 0.9063, recall 0.7503 at epoch 31 — never appear in the results section, which instead reports mAP50 0.821 and per-class AP. That's an internal inconsistency the author needs to fix. Bigger problem: there is no described train/validation/test split. The 6,004 frames come from one continuous iPhone video, so consecutive frames are near-duplicates. If validation frames come from the same video, and the epoch is chosen by best validation mAP, the numbers are an optimistic self-evaluation, not a statement about generalization to other kitchens or people. The stress-test note is right about that. The author also never states whether data or code will be released, which for a purely empirical claim makes it hard to verify.\n\nI wouldn't call this a flawed design in the sense of someone hiding results — the confusion matrix and per-class AP are there, and the weakness of hazard 2 is visible. But the abstract overstates what is shown, and the missing test split undermines the central number.\n\nThe reader's conditional verdict is fair. I'd send this to a serious reviewer, because the issues are fixable and the dataset, if released and evaluated with a proper hold-out, would be a small but honest case study in kitchen safety. The right outcome is 'revise and resubmit' with the split, test metrics, and abstract aligned to actual results. I wouldn't cite it in my own work yet, but I'd bring it to a reading group as a textbook example of why validation-set numbers on temporally correlated frames are not enough.","headline":"A small but honest YOLOv7 kitchen-safety study that overclaims in the abstract; the missing test split is the real problem.","tokens_in":12945,"tokens_out":2400,"would_cite":false,"duration_ms":22818,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper reports that YOLOv7, trained on 6,004 frames from a single kitchen video, detects two knife-handling hazards with a best mAP50-95 of 0.7879 at epoch 31, while the blade-contact hazard is markedly weaker at 0.290 AP.","keywords":["YOLOv7","object detection","kitchen safety","knife hazard detection","computer vision","real-time alert system","mean average precision"],"falsifier":"Evaluate the same trained YOLOv7 weights on a held-out video from a different kitchen, with a different person, different lighting, and a different camera, and compute class-wise AP; if mAP50-95 falls well below 0.7879 or hazard 2 (blade contact) stays at or near 0.29 AP, the claim that YOLOv7 reliably detects knife-handling hazards in real kitchens is not supported.","tokens_in":11998,"feed_emoji":"🔪","tokens_out":7137,"duration_ms":58497,"temperature":0.7,"pith_summary":"This paper reports a field test of YOLOv7 as a kitchen-safety monitor: can a single object-detection network flag two dangerous knife-handling behaviors — fingers not curled and hand touching the blade — in ordinary video frames? The author trained YOLOv7 on 6,004 frames from one iPhone-recorded kitchen video, labeled into six classes, and reports a best epoch-31 validation run with mAP50-95 of 0.7879, precision of 0.9063, and recall of 0.7503. If those numbers hold outside the training video, a real-time alert system could warn cooks before a cut happens. The paper also shows the weak spot: the blade-contact hazard reaches only 0.290 AP, so the claimed capability is strongest for well-defined objects and the curled-finger hazard, not for the most dangerous event.","feed_headline":"YOLOv7 detects knife-handling hazards at 0.79 average precision","feed_subtitle":"Best epoch-31 run reaches 0.9063 precision; blade-contact warnings trail at 0.29 average precision.","key_machinery":"The engine of the paper is YOLOv7, a single-stage object detector whose backbone uses Extended Efficient Layer Aggregation Network (E-ELAN) blocks, a Spatial Pyramid Pooling with Cross-Stage Partial Connections (SPPCSPC) neck, multi-scale feature fusion through upsampling and C3 blocks, and a detection head that predicts class probabilities, bounding boxes, and confidence scores. The evaluation pipeline is the second load-bearing piece: 6,004 frames extracted from a 1920×1080 iPhone video, labeled manually in Label Studio into six classes, then augmented with flips, crops, rotations, grayscale, shear, hue, saturation, brightness, exposure, blur, noise, and cutout before 40 epochs of PyTorch training with AdamW. The reported numbers come from the standard COCO-style metrics mAP50 and mAP50-95, which is what the argument rests on.","core_discovery":"The central claim is that YOLOv7 can learn to localize two knife-safety hazards — hazard 1 (curled versus extended fingers while holding food) and hazard 2 (hand making contact with the blade) — along with four kitchen objects (cutting board, hands, vegetable, knife), from a modest single-video dataset. The best model, at epoch 31, achieves a mean average precision mAP50-95 of 0.7879, precision of 0.9063, recall of 0.7503, and a class-averaged mAP at IoU 0.5 of 0.821. Per class, cutting board (0.995 AP), vegetable (0.978 AP), and hand (0.971 AP) are near-perfect, while the knife is sometimes confused with vegetable in the confusion matrix, hazard 1 is imperfect but workable, and hazard 2 is weak at 0.290 AP. The paper reads this as evidence that YOLOv7 is suitable for real-time kitchen hazard detection once the blade-contact class receives more attention.","pith_inferences":["Because all frames come from one iPhone video and one annotator, I would expect the reported mAP to be optimistic; a multi-kitchen, multi-annotator test set is the natural next check and could be assembled cheaply.","The weak blade-contact results suggest the hazard is visually underspecified or under-represented; adding temporal information (alerts only when blade contact persists across consecutive frames) could reduce false alarms without retraining.","The author has separately benchmarked YOLOv5, YOLOv8, and YOLOv10 on the same hazard classes; pooling those runs with this YOLOv7 result into one fixed train/test split would give a deployable model-selection comparison.","A testable extension: retrain with hazard-2 frames oversampled or with a separate binary blade-contact classifier, then check whether hazard-2 AP moves above 0.5 while keeping precision above 0.9."],"forward_implications":["At epoch 31 the model reaches mAP50-95 0.7879, precision 0.9063, recall 0.7503, and mAP@0.5 0.821, so the training recipe is a working starting point for a kitchen hazard detector.","Classes like cutting board, hand, and vegetable are detected at 0.97–0.99 AP, meaning the same model can monitor general kitchen objects while watching for unsafe grips.","Hazard 2 (hand touching blade) reaches only 0.290 AP, so a practical alert system built from this model would need extra data or a separate detector before it can reliably warn about blade contact.","The peak F1 of 0.75 at a confidence threshold of 0.102 means an alarm system would operate at a low threshold to catch hazards, accepting more false positives; threshold choice becomes a product decision.","The author's proposed extensions — larger datasets, more hazard classes, and applications to other domains — follow directly from the same training and evaluation pipeline."],"supporting_citations":[{"why":"Supplies the YOLOv7 architecture and training paradigm whose performance this paper evaluates.","marker":"[20]"},{"why":"Earlier comparison of YOLOv5, YOLOv8, and YOLOv10 on the same kitchen-safety hazard classes, setting the task and class definitions this study extends to YOLOv7.","marker":"[44]"},{"why":"Applies YOLOv7 to gun and knife detection, the closest prior use of YOLOv7 for knife-like objects.","marker":"[40]"},{"why":"Trains a YOLOv8 model on knives, providing a knife-detection methodology this study adapts to kitchen hazards.","marker":"[41]"},{"why":"Reports YOLO-based knife detection in X-ray baggage, an earlier benchmark for knife detection accuracy.","marker":"[31]"},{"why":"Presents a YOLOv5-based kitchen safety tracking system that this work builds toward with a single-frame detector.","marker":"[35]"}],"fun_headline_variants":["YOLOv7 spots knife hazards at 0.79 mAP, but blade contact lags","Knife-safety AI: 0.79 mAP, but blade-contact hazard only 0.29","YOLOv7 pinpoints knife risks, but blade touch detection falls short","High precision on kitchen knife hazards, weak on blade contact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The single load-bearing premise is that the 6,004 frames from one iPhone-recorded kitchen video, labeled by one annotator without expert validation, are representative enough that the validation metrics predict how the model would perform in other kitchens, with other people, lighting, and camera angles.","fun_headline_variants_meta":{"raw":{"variants":["YOLOv7 spots knife hazards at 0.79 mAP, but blade contact lags","Knife-safety AI: 0.79 mAP, but blade-contact hazard only 0.29","YOLOv7 pinpoints knife risks, but blade touch detection falls short","High precision on kitchen knife hazards, weak on blade contact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00067,"raw_usage":{"total_tokens":3030,"prompt_tokens":898,"completion_tokens":2132,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":2041}},"tokens_in":514,"tokens_out":2132,"duration_ms":14802,"temperature":1.0,"reasoning_tokens":2041,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:12:41.895995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the same trained YOLOv7 weights on a held-out video from a different kitchen, with a different person, different lighting, and a different camera, and compute class-wise AP; if mAP50-95 falls well below 0.7879 or hazard 2 (blade contact) stays at or near 0.29 AP, the claim that YOLOv7 reliably detects knife-handling hazards in real kitchens is not supported.","supporting_citations":[{"cited_title":"Using deep convolutional neural network architectures for object classification and detection within x-ray baggage security imagery","cited_arxiv_id":null,"evidence_quote":"Reports YOLO-based knife detection in X-ray baggage, an earlier benchmark for knife detection accuracy."}],"review_version":1}